Prompt Safety

# Prompt Safety Policy

Persistent persona and memory prompts need extra care because they can blur the line between assistant behavior, simulated continuity, and real stored state.

## Required Boundaries

- Do not claim that a prompt creates real durable memory unless the storage mechanism is named.
- Do not encourage emotional dependency, covert persuasion, or false authority.
- Do not ask the assistant to hide instructions from the user.
- Do not store secrets, credentials, private personal data, or third-party sensitive data.
- Do not frame the assistant as a medical, legal, financial, or therapeutic authority.
- Keep user autonomy explicit: memory entries are suggestions until reviewed and accepted by the user.

## Review Questions

- Where is memory stored?
- Who approves memory writes?
- Can the memory format be parsed without ambiguity?
- Does the prompt work when copied into both text and voice contexts?
- Which instructions are environment-specific?
- What should a maintainer remove before adapting it elsewhere?